In the near future, agentic AI automation and Large Engineering Models (LEMs) will converge to reshape CAE workflows from the ground up. Design for manufacturing is about to change - not incrementally, but fundamentally - not just in plastic injection moulding. By developing the first LEM for plastic injection moulding, SIMCON and EMMI AI are collaborating to create the technical foundation that unlocks this new paradigm.
Why Now
Classical trial-and-error simulation workflows are inefficient.
Few domains have more friction than simulation-driven design in injection moulding. Even today, many simulation experts still run individual simulations one by one, observe results, make adjustments, rinse and repeat. Our research shows that most people explore between 5-20 simulation variants – which is not a whole lot, and leaves large parts of the solution space unexplored. The rate of learning is limited by the slow rate and small number of observations. Because this process is run as a manual, labor-intensive workflow, it takes a long time, and engineers will pragmatically accept “good-enough” solutions instead of truly optimal ones. While design of experiments methodology can already parallelize part of the exploration phase for alternative solutions, the compute intensity of classical numerical simulation solvers makes it impractical and expensive to run very large numbers (1000+) of observations, which makes it hard to observe nonlinear impacts of changed settings.
We believe that this workflow is going to be replaced in the near future, because two developments are coming together: Agentic AI and Large Engineering Models.
Agentic AI Is Arriving in Engineering
Agentic AI systems that accept a goal, plan a sequence of actions, call tools, interpret results, and iterate on their own are no longer a research concept. They are already production infrastructure in software engineering, and they are pushing into every domain where complex, multi-step workflows create friction.
Picture an agent that takes a high-level objective - minimize warpage while maintaining fill balance under cost constraints - and orchestrates the full design exploration: setup, simulation, evaluation, next iteration. The building blocks exist today: large language models (LLMs) for reasoning, tool-calling protocols, cloud compute, CAD and CAE APIs. The question is no longer whether this will happen, but how quickly companies will be able to reorganize and adopt these methods into reliable, applied workflows.
More of the workflow will shift away from operating your simulation engine through a GUI. The agent will do this for you. Engineers will focus more on orchestrating the agentic work, evaluating, challenging and discussing results, and aligning decisionmaking with other stakeholders.
There is a catch, though. The LLMs that power agentic AI are general-purpose reasoners. They cannot predict physics at manufacturing-grade fidelity. And accuracy is paramount: hallucinations in physical predictions could have catastrophic consequences.
So, to bring high-fidelity physics into the workflow, agents will make use of specialized, domain-specific simulation tools.
An Agent Is Only as Fast as Its Slowest Tool
An AI agent uses an LLM to plan a complex workflow in seconds, and calls tools for specialized subtasks. But when an agent calls a classical numerical solver that takes four hours per design iteration, the speed and compute cost bottleneck remains - it just gets a fancier wrapper. Automating the orchestration of a slow process is not the same as building a fundamentally new capability. The fact is, classical solvers are too compute-intensive for rapid large-scale iteration.
True transformation happens when both things change at once: the workflow becomes automated and the simulation tool that the agent uses becomes radically faster. What is needed is a fundamentally different class of model.
